Directing AI
I cannot change the model. The provider owns the weights, the training, and the underlying engineering, and
nothing I do reaches any of that. What I can change is the behavior of the instance I am working in, by what I
have it read and ingest: directives, instructions, constraints, and context. The same out-of-the-box model then
behaves to my standard instead of a generic one. This is not fine-tuning and not a custom model. It is directive
engineering, and the point of this page is that anyone can do it with the right directives in their own instance.
What I cannot touch
The model weights, the training data, the architecture, the inference engine. That is the provider's domain.
No prompt reaches it.
What I can direct
The behavior of my instance: how it reasons, what it refuses, what it verifies, how it writes, and the
standards it holds, all set by directives it reads and ingests.
The directives, and what each one changes
These are the actual directives behind The Conduit, with the observable change each one produced in the instance.
Every effect is something you can open and check.
The organizing law: z = x · y
Instruction: Treat identity times behavior as state. Model parts as
identities with ancestry. AND is z = x · y; weighted AND is
z = x · y².
Effect: Structure stays consistent across modules instead of drifting. The
water health score collapses the moment any one factor fails, rather than averaging a failure away.
The symbolic cycle: INGEST, EXPAND, COLLAPSE, OSCILLATE, INVOKE
Instruction: Run every task through this loop before producing output.
Effect: The instance scopes and checks before it builds, instead of
jumping to an answer. The agent loop in the project follows it step for step.
No claim without proof
Instruction: A claim is added only when it is tested and proven. Until
then it stays a hypothesis and is labeled as one.
Effect: Tests are run and output is shown before anything is called done.
Unverified work is marked unverified, not presented as working.
Synthesis is not recovery
Instruction: Detail that is added is synthesized, not recovered. Say so
plainly.
Effect: The tools state that they add plausible detail and do not recover
lost data. A scoring function is labeled a heuristic, not a judge.
Simulation only, never fabricate operational data
Instruction: The water system is a simulation. No real data. Keep the
disclaimer visible.
Effect: The 3D level-of-detail reveals only data that actually exists and
stops where the data ends, rather than inventing infrastructure to fill the view.
Authority decides, software recommends
Instruction: The software recommends GO or NO-GO. A named human authority
always makes the final call.
Effect: The go-live gate was built as advisory, with a documented human
override, not an automated lock.
A style directive: no em-dashes
Instruction: Do not use em-dashes.
Effect: Output style changes immediately and stays changed. A small,
easily checked proof that an instance instruction shifts behavior at once.
Anti-hallucination reasoning
Instruction: Reason with truth tables, logic gates, determination graphs,
and decision trees. Bound the recursion.
Effect: Steps stay explicit and checkable, and expansion stays bounded
instead of wandering into unrelated territory.